Add _redact_secrets() in builtin_tools/security.py and apply it at every
commit_memory call site before content reaches the memories table.
Patterns scrubbed (replaced with [REDACTED]):
- sk-[A-Za-z0-9_-]{20,} OpenAI/Anthropic keys (sk-, sk-ant-, sk-proj-)
- ghp_[A-Za-z0-9]{36} GitHub classic PAT
- ghs_[A-Za-z0-9]{36} GitHub server-to-server token
- github_pat_[A-Za-z0-9_]{82} GitHub fine-grained PAT
- AKIA[0-9A-Z]{16} AWS access key ID
- key/token/secret/password/api_key=<40+ chars> Generic contextual (value replaced,
keyword preserved: "api_key=[REDACTED]" not "[REDACTED]")
Call sites wired:
- builtin_tools/memory.py::commit_memory() — LangChain tool (LangGraph path)
- a2a_tools.py::tool_commit_memory() — MCP server path
- executor_helpers.py::commit_memory() — CLI/SDK executor path
Implementation guarantees:
- Pure function (no side effects, no I/O)
- Idempotent: [REDACTED] does not match any pattern
- No false positives on normal prose (all patterns require ≥20-char prefix
or ≥40-char value after known keyword)
Tests (36 passing):
- Per-pattern unit tests for all 6 secret types
- Idempotency tests
- Normal prose non-regression tests
- Integration: a2a_tools.tool_commit_memory scrubs ghp_ tokens before HTTP POST
- Integration: executor_helpers.commit_memory scrubs AWS keys and OpenAI keys
- Source inspection: memory.py imports and applies _redact_secrets before
build_awareness_client() (i.e. before any storage operation)
conftest.py updated to load the real builtin_tools/security.py so that
executor_helpers and a2a_tools can import _redact_secrets during test collection.
Closes#834
Sub-issue of #725
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- WorkspaceNode.eject.test.tsx: add draggable/selectable/deletable to
NodeProps render call (TS2739); add `as WorkspaceNodeData` cast on
makeNodeData return to silence Partial<> spread widening (TS2322)
The cherry-picked fix/canvas-test-fixture-budgetlimit commit (9e0aa61)
also lands here — it resolves latent test-fixture drift in 7 test files
that the incremental tsc cache had masked on main but that became visible
once the new WorkspaceNode.eject.test.tsx file invalidated the cache.
tsc --noEmit: 0 errors | npm test: 726 passed | npm run build: clean
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The budget PR (#541) added budgetLimit: number | null as a required field
on WorkspaceNodeData and budget_limit: number | null on WorkspaceData.
Seven test fixture factories were not updated, causing tsc --noEmit to
produce 34 TS2322/TS2345 errors (runtime tests still passed because
Vitest transpiles via esbuild which strips types).
Fixes:
- canvas-events.test.ts: makeNode factory +budgetLimit: null
- canvas-events-pan.test.ts: makeNode factory +budgetLimit: null
- canvas-capabilities.test.ts: makeNodeData factory +budgetLimit: null
- canvas-topology.test.ts: makeWS factory +budget_limit: null
- canvas.test.ts: makeWS factory +budget_limit: null; two inline
summarizeWorkspaceCapabilities args +budgetLimit: null; context-menu
fixture +budgetLimit: null
- ProvisioningTimeout.test.tsx: makeWS factory +budget_limit: null
Also fixes 3 TS2348 errors in AuthGate.test.tsx: newer Vitest type defs
resolve ReturnType<typeof vi.fn> to Mock<Procedure|Constructable> which
TypeScript no longer considers directly callable in a vi.mock factory.
Fix: intersect the mock variables with a plain function type so both the
call expression and the mock API (mockReturnValue etc.) type-check.
tsc --noEmit: 0 errors. npm test: 722/722.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
PR#842 merged the docs/opencode.json to main with the correct MCP URL path.
PR#840 branch had an older version — sync to main's content to resolve conflict.
PR#842 merged the docs/opencode.json to main with the correct MCP URL path.
PR#840 branch had an older version — sync to main's content to resolve conflict.
Security Auditor pre-merge conditions for PR#840:
C5: toolCommitMemory passes content directly to DB insert without secret
redaction. Gap is tracked to #838 (platform-wide _redactSecrets pass).
Adds inline TODO(#838) comment at the insert site so the gap is visible
in-code, not only in the issue tracker.
C6: toolDelegateTask sets X-Workspace-ID but no bearer token on the
outbound A2A call. The /workspaces/:id/a2a route is intentionally outside
WorkspaceAuth (by design in router.go). CanCommunicate is enforced before
the request is constructed, and callerID was authenticated by WorkspaceAuth
on the MCP bridge entry point. Documents this trust assumption at the call
site.
The inline JSON example still showed the bare ${MOLECULE_MCP_URL} without
the /workspaces/${WORKSPACE_ID}/mcp path. Updated to match opencode.json fix
in previous commit (bf80f15). Added WORKSPACE_ID to the env section.
Tables: Slack has no table syntax. Converter now detects markdown tables
and renders them as monospace code blocks with aligned columns.
Dividers: replaced unicode em-dash (caused encoding artifacts) with
plain ASCII dashes.
Strikethrough: ~~text~~ converts to ~text~ (Slack native).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The function was defined on a feature branch, referenced by manager.go
and slack_test.go, but never made it to main after the rebase. This
caused go build to fail with 'undefined: FetchChannelHistory', which
Docker masked by using a cached binary from the last successful build.
That cached binary had neither the mrkdwn blocks nor the Level 3
context injection — explaining why Slack messages showed raw markdown
despite the source having the converter.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- EjectIcon now accepts React.SVGProps<SVGSVGElement> so aria-hidden can be passed
- Eject button: aria-label and title both use `Extract ${data.name} from team`
(previously title was static 'Extract from team'; aria-label was absent)
- <EjectIcon aria-hidden="true"> prevents assistive tech from double-announcing
the icon content inside the already-labelled button
- Added WorkspaceNode.eject.test.tsx (4 tests) covering aria-label, title,
label==title invariant, and aria-hidden on the SVG
Closes#854
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace denylist approach with strict allowlist: only PATH, HOME, LANG,
PYTHONPATH, WORKSPACE_ID, WORKSPACE_NAME, PLATFORM_URL (and a small set
of locale/Python runtime vars) pass through to agent-executed code. Every
other env var — including ANTHROPIC_API_KEY, GH_TOKEN, DATABASE_URL,
REDIS_URL, *_SECRET, *_PASSWORD — is stripped from os.environ for the
duration of SafeLocalPythonExecutor.__call__ and restored on exit.
- make_safe_env() is a pure read (never mutates os.environ)
- _ENV_PATCH_LOCK serialises concurrent calls for thread safety
- os.environ fully restored even on exception (try/finally)
- 38 unit tests covering all secret categories, thread safety, import
restrictions, and env-restore guarantees
Closes#826
Sub-issue of #804
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Slack's chat.postMessage renders the text field as plain text when
username override is used. Switching to blocks with type=mrkdwn
forces rich formatting (bold, links, code, dividers).
Also restores FetchWorkspaceChannelContext that was lost in rebase.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Both were lost during the PR #844 rebase — the converter was in the
source but the binary couldn't compile because FetchWorkspaceChannelContext
was missing from manager.go (interface mismatch). Previous deploys
silently used the cached old binary without the converter.
Also removed unused 'log' import that blocked compilation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The budget PR (#541) added budgetLimit: number | null as a required field
on WorkspaceNodeData and budget_limit: number | null on WorkspaceData.
Seven test fixture factories were not updated, causing tsc --noEmit to
produce 34 TS2322/TS2345 errors (runtime tests still passed because
Vitest transpiles via esbuild which strips types).
Fixes:
- canvas-events.test.ts: makeNode factory +budgetLimit: null
- canvas-events-pan.test.ts: makeNode factory +budgetLimit: null
- canvas-capabilities.test.ts: makeNodeData factory +budgetLimit: null
- canvas-topology.test.ts: makeWS factory +budget_limit: null
- canvas.test.ts: makeWS factory +budget_limit: null; two inline
summarizeWorkspaceCapabilities args +budgetLimit: null; context-menu
fixture +budgetLimit: null
- ProvisioningTimeout.test.tsx: makeWS factory +budget_limit: null
Also fixes 3 TS2348 errors in AuthGate.test.tsx: newer Vitest type defs
resolve ReturnType<typeof vi.fn> to Mock<Procedure|Constructable> which
TypeScript no longer considers directly callable in a vi.mock factory.
Fix: intersect the mock variables with a plain function type so both the
call expression and the mock API (mockReturnValue etc.) type-check.
tsc --noEmit: 0 errors. npm test: 722/722.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Agents output standard Markdown (Claude Code default) but Slack uses
its own mrkdwn format. Without conversion:
**bold** shows as literal **bold**
### heading shows as literal ###
[text](url) shows as raw markdown link
Converter handles:
**bold** → *bold* (Slack bold is single asterisk)
### heading → *heading* (bold text, no headings in Slack)
[text](url) → <url|text> (Slack link format)
--- → ——— (visual separator)
`code` and ```blocks``` pass through unchanged
6 new tests: bold, heading, link, hr, code block, mixed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two WCAG violations in the Danger Zone delete flow:
1. WCAG 4.1.3 (Status Messages): the confirmation UI that appears when
the user clicks "Delete Workspace" had no ARIA live region, so screen
readers never announced the confirmation prompt. Adding role="alert"
to the confirmation container makes it an implicit assertive live
region that is announced immediately.
2. WCAG 2.4.3 (Focus Order): pressing Cancel left focus wherever the
browser placed it (often body). Keyboard users had to re-navigate to
find the Delete Workspace button. The Cancel handler now calls
deleteButtonRef.current?.focus() to return focus to the trigger
button, matching the expected modal/disclosure focus-management pattern.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
WCAG 1.3.1 / 4.1.3: the onboarding card had no landmark role and no
live region, so screen readers had no way to know the card exists or
that the step changed.
- Add role="complementary" aria-label="Onboarding guide" to the card
container so it appears as a named landmark in assistive technology.
- Add a role="status" aria-live="polite" aria-atomic="true" sr-only div
that holds the current step label. When the step state changes React
updates the div content, which the live region broadcasts to the AT
without pulling focus away from the user's current position.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
NVDA and other screen readers ignore the title attribute on interactive
elements and non-interactive divs. Add aria-label alongside title on:
- Stop All button (dynamic label reflects active task count)
- Restart All button (dynamic label reflects pending workspace count)
- StatusPill component (online/offline/failed/provisioning counts)
- WsStatusPill component (connected/connecting/disconnected variants)
Inner dot and text spans get aria-hidden="true" so the screen reader
reads the single aria-label rather than individual child nodes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
WCAG 1.3.1 / 4.1.3: the error div that appears after a failed workspace
deploy or blank-workspace create had no ARIA live region, so screen
readers never announced it. Adding role="alert" makes the message an
implicit aria-live="assertive" region so assistive technology surfaces
the error immediately without requiring the user to navigate to it.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wire WCAG 1.3.1 label associations: 6 bare <label>+control pairs in
ConfigTab (Description, Tier, Runtime, Effort, Task Budget, Backend) now
use stable useId() IDs with matching htmlFor/id. Field helper in
DetailsTab updated to generate its own fieldId via useId() and inject it
into the child element via cloneElement, so every Name/Role/Tier field in
edit mode is correctly associated without requiring call-site changes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a cron fires, the scheduler now fetches the last 10 messages from
the workspace's Slack channel via conversations.history and prepends them
to the cron prompt as '[Slack channel context — recent team messages]'.
This gives each agent ambient awareness of what peers are doing:
- Backend sees Frontend posted 'PR #840 ready for review' → can check
- Security Auditor sees Backend posted 'new endpoint added' → plans review
- PM sees all engineering activity → better synthesis in rollup
Implementation:
- slack.go: FetchChannelHistory() calls conversations.history, filters
bot's own messages, returns last N as SlackHistoryMessage structs
- manager.go: FetchWorkspaceChannelContext() looks up the workspace's
Slack config, fetches history, formats as readable context block
- scheduler.go: ChannelBroadcaster interface extended with
FetchWorkspaceChannelContext; fireSchedule injects context before
the cron prompt (prepended, not appended, so the agent sees team
context BEFORE its task instructions)
Best-effort: if Slack API fails or workspace has no channels, the
prompt is unchanged. Truncated to 200 chars per message, 10 messages
max to keep prompt overhead bounded.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Code review findings addressed:
Critical:
1. Bot echo loop: add bot_id + subtype='bot_message' check in ParseWebhook
to prevent outbound auto-posts from triggering inbound → infinite loop
2. Connection leak: close resp.Body immediately after reading instead of
defer inside loop (was holding N connections open for N chunks)
3. Cancelled context: auto-post goroutine now uses context.Background()
with 30s timeout instead of inheriting fireCtx (which gets cancelled
by deferred cancel() when fireSchedule returns)
4. Slug validation: regex ^[a-zA-Z0-9 _-]+$ rejects path traversal and
special chars in [slug] routing
Improvements:
5. Shared HTTP client (slackHTTPClient) for connection pooling instead of
per-request &http.Client{}
6. Rune-safe truncation in BroadcastToWorkspaceChannels for CJK/emoji
7. Log async HandleInbound errors instead of silently discarding
8. url_verification challenge properly returned (c.JSON with challenge)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Humans type [backend] what's #800? in a shared #mol-engineering channel
and the message routes specifically to Backend Engineer's workspace.
Matching logic (case-insensitive):
[pm] → PM
[backend] → Backend Engineer
[dev-lead] → Dev Lead
[security] → Security Auditor (prefix match on 'security-auditor')
Unknown slugs return the available agent list for that channel so the
user knows what slugs are valid.
Messages without a [slug] prefix route to the first matching workspace
(backward compat with Level 2).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Level 1 — Auto-post cron output to Slack:
- scheduler.go: captures A2A response body, extracts agent text via
extractResponseSummary(), broadcasts to workspace's configured Slack
channels on successful non-empty cron completions
- manager.go: adds BroadcastToWorkspaceChannels() — fans out to all
enabled channels for a workspace (engineering+firehose for eng agents,
research+firehose for research agents, etc.)
- main.go: wires scheduler → channel manager via SetChannels()
- Truncates output to 500 chars for Slack readability
Level 2 — Inbound Slack messages route to workspaces:
Already implemented by the existing webhook handler (POST /webhooks/slack)
+ the ParseWebhook method in slack.go which handles both Events API JSON
payloads and slash command form-encoded payloads. Needs Slack App Events
API URL configured to: https://<platform-host>/webhooks/slack
Also in this commit:
- slack.go: dual-mode adapter (bot_token + webhook fallback)
- 031 migration: pgvector guard wraps entire DO block
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Slack adapter: adds chat.postMessage mode alongside legacy webhooks.
When bot_token is configured, uses chat:write.customize for per-agent
display name + emoji on every message. Each of the 15 active agents
posts with a distinct identity (PM 💼, Backend ⚙️, etc.).
5 channels configured:
#mol-engineering — PM, Dev Lead, Frontend, Backend, QA, Security, UIUX, Docs
#mol-research — Research Lead, Market Analyst, Tech Researcher, Competitive Intel
#mol-ops — DevOps, Triage, Offensive Security
#mol-ceo-feed — PM synthesized rollup (CEO-facing)
#mol-firehose — all agents (raw feed)
Tested live: 5 test messages across 4 channels, all ok=true.
pgvector migration: moved ALTER TABLE + CREATE INDEX inside the DO
block so the entire migration is skipped when pgvector extension is
unavailable (was crashing platform on restart — the guard caught
CREATE EXTENSION but execution continued to ALTER TABLE which used
the non-existent vector type).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces the racy SELECT-then-Stop two-step in HibernateWorkspace with a
three-step atomic pattern that eliminates the TOCTOU window (SAFE-819):
1. Atomic claim: single UPDATE WHERE id=$1
AND status IN ('online','degraded')
AND active_tasks = 0
— rowsAffected=0 means another caller already claimed it or tasks
arrived; we abort immediately without calling Stop.
2. provisioner.Stop: safe because status='hibernating' blocks new task
routing between step 1 and step 2 (no new task can be dispatched).
3. Final UPDATE to 'hibernated': records the completed hibernation.
Also adds stopFnOverride func(ctx, id) to WorkspaceHandler (always nil in
production) so tests can count Stop calls without a running Docker daemon.
Tests added/updated (13 total across 2 files):
- TestHibernateWorkspace_ActiveTasksNotHibernated
- TestHibernateWorkspace_AlreadyHibernatingNotHibernated
- TestHibernateWorkspace_SuccessPath
- TestHibernateWorkspace_ConcurrentOnlyOneStop
- TestHibernateWorkspace_DBErrorOnClaim
- Updated 3 existing HibernateWorkspace tests + 1 HTTP handler test
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add role="separator" + aria-valuenow/min/max/orientation + tabIndex={0}
to make the resize handle focusable and discoverable by screen readers
(WAI-ARIA slider pattern). Add onKeyDown handler: ArrowLeft/Right moves
by 16px, Home/End snaps to min/max. Persist width to localStorage on
keyboard resize, matching the existing mouse behaviour.
Focus ring uses focus-visible:ring-2 to avoid showing on mouse click.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously loadMessagesFromDB swallowed all errors and returned [] — a
network failure was indistinguishable from an empty history, so the user
had no way to know loading failed. Now the function returns
{ messages, error } and the MyChatPanel renders a role="alert" banner
with the error message and a Retry button when messages are empty and
a load error occurred.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace title attribute (not read by screen readers for truncated text)
with aria-label, add role="status" so live regions announce the error,
and raise text color from text-amber-300/60 (~2.1:1) to text-amber-400
(~10.6:1) to meet WCAG AA contrast (4.5:1 minimum).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add bodyId derived from entry.key, attach aria-controls={bodyId} to the
toggle button, and add id={bodyId} role="region" aria-label to the
collapsible body div. Screen readers can now announce the expand/collapse
relationship between the button and the region it controls (WCAG 4.1.2).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The migration SQL is read as raw SQL (not through Go fmt.Sprintf),
so %% is two parameters, not an escaped percent. Postgres RAISE
uses single % for parameter substitution.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The ALTER TABLE and CREATE INDEX referenced vector(1536) outside the
exception-handling DO block, so when pgvector wasn't installed they
crashed the migration runner — blocking ALL E2E runs on main.
Fix: move all DDL inside the single DO block so the EXCEPTION handler
catches any pgvector-related failure and skips the entire migration.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
